ShelfWright

1. One-liner

Rewrites every product record your catalog is missing, so AI shopping agents can actually recommend what you sell.

2. Trend signal — why now?

A new distribution surface appeared in eighteen months and most merchants are structurally locked out of it — not by permission, but by data.

The surface is real and it converts. Adobe Analytics clocked 393% year-over-year growth in AI traffic to U.S. retail in Q1 2026, and — the number that matters — AI-referred shoppers now convert 42% better than non-AI traffic, reversing a 38% under-performance a year earlier. That’s an ~80-point swing in twelve months. ChatGPT sits at 900M weekly actives running roughly 50M shopping queries a day. Salesforce attributed ~20% of 2025 holiday orders ($262B) to agent-driven flows.

The pipe is now free, which is exactly what makes this interesting. Shopify flipped Agentic Storefronts on by default for eligible US merchants in March 2026 — opt-out, not opt-in — pushing millions of stores into ChatGPT, Copilot, Google AI Mode and Gemini automatically. Every “connect your store to ChatGPT” startup just got its core feature commoditized to zero overnight.

But connection is not visibility. The plumbing arrived; the catalogs didn’t. CommerceShop audited 305 SMB ecommerce brands ($2.2B combined revenue) across Shopify, Shopify Plus and WooCommerce and found less than 1% had what’s required for full AI visibility. Product Schema adoption: 63% on Shopify Plus, 54.6% on Shopify, 24% on WooCommerce. Ninety-nine of the brands had FAQ content that “ChatGPT, Gemini, and Perplexity literally cannot read” because the JSON-LD was malformed. Their conclusion is the thesis of this idea: “Revenue does not buy AI visibility. Implementation does.”

And the failure modes are specific, mechanical, and fixable. The single most common ACP eligibility failure is missing GTINs at the variant level. Then: stale variant-level inventory, policy URLs that 404 (the most common checkout-eligibility rejection), prices not in ISO 4217 minor units (7999, not 79.99), non-unique product IDs, marketing prose where factual descriptions belong, and weak variant titles — “Blue / Size 8” where the agent needs “Cobalt Blue / Women’s US Size 8 / Wide Width.” Merchants with complete attributes (GTIN, brand, material, color, size, use-case) surface in materially more AI queries; one analysis puts near-complete catalogs at 3–4× the AI recommendation visibility of sparse ones.

Worst part for the merchant: submitting a bad catalog isn’t neutral, it’s worse than not submitting — thin records get down-ranked and drag the whole feed’s quality score.

Money is moving. Agentic commerce seed rounds closed through 2026 — ReFiBuy $13.6M (May 2026), ShopAgentic €1.9M pre-seed (June 2026), Merchantee €1.8M, Nudge $1.1M — concentrated in payments/identity infrastructure, not in SMB catalog remediation. Meanwhile AEO agencies charge $2,500–$25,000/mo with one-time audits at $5,000–$15,000, which prices out exactly the 500–20,000-SKU brand that has the problem worst.

Provenance:
  - Signal 1 (Demand): 305-brand SMB audit finds <1% AI-search-ready; WooCommerce Product Schema adoption just 24% — https://www.thecommerceshop.com/ebook/state-of-ai-search-readiness-smb-ecommerce-2026/ — 2026
  - Signal 2 (Feasibility): Shopify activates Agentic Storefronts by default (March 2026) + OpenAI ACP published as an open feed spec with concrete required fields and 15-minute refresh — https://developers.openai.com/commerce/specs/spec , https://novadata.io/resources/news/shopify-agentic-storefronts-chatgpt-default-2026 — March 2026
  - Signal 3 (Economic): AI retail traffic +393% YoY and converting 42% better than non-AI (Adobe, Q1 2026); agentic-commerce seed capital flowing (ReFiBuy $13.6M, May 2026) while AEO agencies charge $2.5K–$25K/mo — https://elogic.co/blog/chatgpt-commerce-statistics/ , https://retailtechinnovationhub.com/home/2026/5/5/agentic-commerce-specialist-refibuy-raises-136-million-seed-round-led-by-newroad-capital-partners — 2026
  Category: Platform shift

3. The opportunity

Three surfaces — OpenAI ACP, Google/Gemini via Merchant Center + UCP, Perplexity’s merchant program — each want the same catalog described differently. Perplexity takes a Google Shopping CSV over SFTP and treats GTIN as mandatory. Google’s UCP path wants a /.well-known/ucp profile and ≥50 approved GMC offers. OpenAI wants JSONL with variant-level identifiers, minor-unit pricing, and three live policy URLs. One catalog, three dialects, all drifting.

The existing market splits into two halves and neither serves the middle:

Free validators — Rankfender’s ChatGPT Shopping Feed Checker, UCPtools’ ACP Checker and Feed Analyzer. They tell you what’s broken. They hand a merchant a list of 4,000 errors and a shrug. Diagnosis is already free and commoditized; nobody should try to sell it.

Feed managers — DataFeedWatch (~$64–$239/mo), Channable (from ~€49/mo), Feedonomics and Productsup (quote-only, sales-led, enterprise). These are transport layers: excellent at mapping a field you already have to a channel that wants it, renaming colour to color, splitting a feed by region. They do not manufacture the attribute that was never in your PIM. If material is empty on 3,100 SKUs, DataFeedWatch will faithfully transmit 3,100 empty material fields.

The gap is the fix itself. Between “here are your errors” (free) and “$8K/mo agency retainer” (unaffordable) there is nothing that takes a 6,000-SKU catalog and writes the missing data — infers material and use-case from images and existing copy, generates spec-grade variant titles, assigns Google Product Category, decides GTIN-vs-brand+MPN correctly for private-label goods, catches the 404’d returns policy — then keeps it correct as the specs mutate.

That’s the 10×. An agency does this with junior staff at $8K/mo for a few hundred SKUs. A vision-plus-language model does the same enrichment per-SKU for fractions of a cent, which is what makes a $149–$599/mo price point possible on a job that was quoted at five figures eighteen months ago.

4. Target market

• Primary customer: Founder or head of ecommerce at a DTC/niche brand doing $1M–$30M GMV with 500–20,000 SKUs on Shopify, Shopify Plus, WooCommerce or BigCommerce. Apparel, home goods, outdoor gear, specialty food, auto parts, beauty. Typically 2–15 people with nobody who owns “product data” as a job.

• Why they buy: They know AI traffic is the one channel still growing and converting above baseline while paid social CAC keeps climbing. They’ve heard they’re “already connected.” They can’t tell whether they’re actually being recommended, and when they run a free checker they get an error list they have no labor to act on. The merchant question is blunt and recurring — a private-label skincare seller on the Shopify forums asking whether self-manufactured brands “need GTINs to be eligible for Google AI Overviews product pulls,” where the answer turns out to be nuanced (brand + MPN is fine for D2C, GTIN matters for marketplaces) in a way no merchant should have to research per-attribute.

• Rough TAM reasoning: ~4.17M live WooCommerce stores and ~4.7–5.5M Shopify merchants. Cut hard to the band with real catalog depth and revenue to defend — call it 250K–400K brands globally in the 500+ SKU, $1M+ GMV range. At a $299 blended ACV/mo, capturing 0.5% of the low end is ~$45M ARR. This idea needs 300 customers, not 30,000. The market is not the constraint.

• Why now for them: Shopify’s March 2026 opt-out default means their catalog is already being read by agents today, in whatever state it’s in. Before March, bad data meant absence. Now bad data means active down-ranking and a damaged feed quality score. The clock started without them.

5. Product sketch (MVP)

• Connect and score. OAuth into Shopify/Woo/BigCommerce, pull the catalog, return a per-surface readiness score (ACP / Google-UCP / Perplexity) with SKU-level breakdown of exactly which records fail which rule and what that blocks — search eligibility vs. checkout eligibility.
• Bulk attribute enrichment. Infer and draft the missing fields — material, color, dimensions, use-case, Google Product Category, spec-grade variant titles — from existing images, descriptions and supplier docs. Every value is a proposal, never an auto-publish.
• Identifier resolution. Per-SKU call on GTIN vs. brand+MPN, flagging which SKUs genuinely need a GS1 barcode versus which are correctly exempt as private-label — the single most common eligibility failure, handled as a decision rather than an error message.
• Merchant-in-the-loop review queue. Approve, edit or reject in bulk by category. Nothing reaches the storefront unreviewed. Accepted values write back to the store as the source of truth.
• Policy and structural checks. Verify return/privacy/ToS URLs resolve and aren’t login-gated, price formatting is in minor units, variant IDs are stable and unique, inventory is variant-level, Product/Offer schema is valid JSON-LD.
• Drift monitor. Re-scan on a schedule; alert when new SKUs land incomplete, when a policy URL breaks, or when a surface changes its spec — with a diff of what to fix.
• Surface-specific export. Generate the compliant artifact per channel: ACP JSONL, Google Shopping CSV for Perplexity’s SFTP, GMC-ready output.

6. AI angle — what’s load-bearing

Delete the AI and this product is a validator — which already exists, free, from two vendors.

The load-bearing work is manufacturing attributes that were never recorded. Reading a product photo to determine that a jacket shell is ripstop nylon. Turning three sentences of brand voice into a factual, agent-parseable description without losing accuracy. Converting “Blue / Size 8” into “Cobalt Blue / Women’s US Size 8 / Wide Width.” Classifying 6,000 SKUs into Google Product Categories. Extracting specs from a supplier PDF nobody has opened since 2023.

This is per-SKU inference at catalog scale — thousands of small vision-and-language judgments that were previously junior-analyst labor billed at agency rates. That’s precisely the shape of work the persona filter asks for: a task that cost $8K/mo of human attention collapsing to a job that runs in an hour for a few dollars of inference. The economics of the product are the economics of the model.

Second-order: enrichment quality compounds. Which drafted attributes merchants accept, edit, or reject is direct labeled feedback, per category, that sharpens the next merchant’s suggestions.

7. Localization angle (if any)

N/A — this is a global play, English-first. The buyer constraint is catalog depth and platform, not geography. Worth noting the surfaces are US-gated today (ChatGPT in-chat checkout is US merchants and US users only; OpenAI’s self-serve merchant portal is in early access with Europe “later in 2026”), so v1 targets US and UK/EU-exporting brands. The natural localization expansion is later and mechanical: multilingual attribute generation for brands selling into DE/FR/ES, where an agent must parse the catalog in-language — that’s an ACV expansion lever, not a launch wedge.

8. Business model — path to $1M–$5M ARR

• Pricing: SKU-banded SaaS.

  • Starter — $149/mo: up to 1,000 SKUs, one surface, monthly re-scan.
  • Growth — $349/mo: up to 7,500 SKUs, all three surfaces, weekly drift monitoring.
  • Pro — $799/mo: up to 25,000 SKUs, daily monitoring, multi-store, API.
  • One-time remediation sprint — $1,500–$4,000: full first-pass enrichment, delivered reviewed. This is the door-opener and it prices against $5K–$15K agency audits.

• ACV: ~$3,600 blended ($299/mo average), plus roughly $900 of first-year one-time sprint revenue on ~40% of accounts.

• Rough math to $1M ARR: 280 customers × $299/mo × 12 = $1.0M. Achievable inside 18 months with the channels in §9.

• Rough math to $5M ARR: ~1,150 customers at a $360 blended ACV/mo — requires the agency/3PL reseller channel carrying ~40% of new logos, plus expansion into multi-store and multilingual tiers. Realistically 30–36 months, which is at the outer edge of the persona’s window. $1M–$2M is the honest near-term target; $5M needs the partner channel to work.

• Expansion path: SKU-band upgrades as catalogs grow → additional surfaces as new agent platforms publish specs (each new protocol is a paid module, not a rebuild) → multilingual enrichment → multi-store/agency seats.

• Margin note: enrichment inference is the real COGS and it is front-loaded at onboarding, not recurring. A 7,500-SKU first pass costs single-digit dollars against a $349/mo subscription; steady-state drift monitoring is near-free. Gross margin should sit above 85% after the first month of each account.

9. Go-to-market wedge — first 100 customers

• Free public scanner as the entire top of funnel. Merchant enters a store URL; get back a real per-surface readiness score and the top 10 failing SKUs by name. Free tools already prove demand for this diagnosis — the difference is that the fix button is right there. Every scan is a qualified lead with the problem already quantified in their own catalog.
• Cold outreach with the audit pre-run — the primary motion. Scrape 3,000 Shopify/Woo brands in the 500+ SKU band from BuiltWith and public store directories, run the scanner on each before contacting them, and send a 90-second Loom showing their three worst SKUs and what an agent sees instead of their product. This is the CommerceShop finding as a sales weapon: fewer than 1% will pass, so nearly every email carries a real, specific, embarrassing defect. Expect 4–7% reply on a personalized-artifact cold email, 20% of replies to demo, ~25% of demos to close a remediation sprint → ~15–25 customers from the first 3,000.
• Agency and 3PL reseller channel. AEO/ecommerce agencies charging $2,500–$25,000/mo cannot profitably do SKU-level remediation with human labor. Sign 10 agencies as white-label resellers at 25% recurring margin; each brings 3–8 client catalogs. This is the only path that gets past $2M ARR.
• Ride the specification churn. Every time OpenAI, Google or Perplexity revises a feed spec, publish a same-week teardown naming the exact fields that changed and who breaks. Merchants and agencies searching that error string in the following fortnight land on the scanner. Spec volatility is permanent here — the 30% citation-instability finding says this churn isn’t settling down soon.
• Platform app stores. Shopify and BigCommerce app listings, plus a WooCommerce plugin. WooCommerce is the sharpest target — 24% schema adoption versus Shopify’s 54.6%, no auto-syndication, and merchants who must submit manually. Worst-served, most acute pain, weakest incumbent coverage.

10. Build complexity — justification

Low. Everything is off-the-shelf: platform OAuth and product APIs, published feed specs from OpenAI/Google/Perplexity, standard vision-and-language model calls for enrichment, a review queue, a scheduler. No custom models, no novel infrastructure, no proprietary dataset needed at launch.

The genuine work is unglamorous rather than hard: encoding each surface’s validation rules faithfully, building a bulk-review UI that makes approving 4,000 suggested attributes tolerable rather than maddening, and safe write-back to live storefronts. Solo builder: 7–9 weeks to a paid v1 covering Shopify + one surface; add 4–6 weeks for Woo/BigCommerce and the remaining two surfaces. The free scanner can ship in two weeks and start collecting leads while the fix engine is built.

11. Gating checklist

Gate	Pass?	Note
Legal in target market	✅	Merchant-authorized access to their own catalog via official OAuth. Published, public feed specs. No scraping of competitor data.
Ethical — no harm / dark patterns	✅	Merchant-in-the-loop by design; nothing publishes unreviewed. Explicit guardrail: enrichment must be inferred from the merchant’s own assets, never invented. Fabricated specs would harm shoppers and the merchant — this is the product’s central ethical constraint, addressed in §13.
Market exists (evidence above)	✅	305-brand audit (<1% ready), free validators with traction, agencies billing $2.5K–$25K/mo for the manual version.
1–5 person team can build this	✅	Solo to pair, 7–9 weeks to paid v1.
Launchable with <$50K / ₹40L	✅	Inference, hosting, a scraped prospect list. Well under $10K to first revenue.

12. Feasibility score

Axis	Weight	Score	Notes
Problem intensity	20	15/20	Real and quantified — <1% of audited brands ready, bad data now actively down-ranks. But it’s a growth pain, not a compliance deadline. Nobody gets fined. A merchant can ignore this for another quarter and only lose upside, which caps urgency below the hair-on-fire band.
Demand evidence	15	13/15	Multiple independent hard signals: 305-brand audit with methodology, Adobe traffic/conversion data, free validators already built (someone saw the demand), agencies charging five figures for the manual version, funded startups adjacent. A skeptic nods.
Build feasibility	15	13/15	Off-the-shelf throughout. Solo v1 in 7–9 weeks. Bulk-review UX is the only genuinely fiddly piece.
Distribution clarity	15	12/15	Free scanner → pre-run audit → Loom is concrete, named, and the conversion math is defensible because virtually every prospect fails. Docked for reliance on cold email reply rates and an unproven agency channel.
Revenue mechanics	15	11/15	Pricing is benchmarked between free tools and $2.5K/mo agencies, margins are strong, $1M path needs only 280 customers. Docked because $5M needs the reseller channel to work and churn is a real unknown — a merchant who gets “fixed” may not see why they keep paying.
Time to first revenue	10	8/10	Remediation sprints can be sold before the recurring product is finished; realistically 6–8 weeks to first dollar.
Defensibility	10	2/10	The weak axis and I won’t dress it up. Feed managers (DataFeedWatch, Channable, Feedonomics) can bolt on AI enrichment; Shopify itself could ship catalog suggestions natively and vaporize the category overnight. Only real moats are accumulated accept/reject data by category and workflow lock-in as the system of record for catalog quality — both take 12+ months to matter. Month 3: fully copyable. Month 12: a niche brand and a data flywheel, nothing more.
Total	100	74/100

13. Qualitative modifiers

Founder-fit tags

technical-heavy · content-heavy

Technical for the platform integrations, enrichment pipeline and per-surface rule engines. Content-heavy because the spec-churn teardown strategy in §9 is a primary acquisition channel and needs someone who will actually publish within a week of every spec revision.

Key assumptions to validate (3–5)

1. Assumption: Merchants will pay recurring, not just for a one-time cleanup — i.e. catalog drift is continuous enough to justify a subscription. How to test: run the free scanner against 200 stores twice, six weeks apart. Measure how many new incomplete SKUs and newly-broken policy URLs appear per store. If the median store adds fewer than ~20 defects in six weeks, this is a services business with a $2K one-time fee, not SaaS — reprice accordingly.
2. Assumption: AI-surface visibility gains are attributable enough that a merchant can see ROI and renew. How to test: instrument 10 pilot accounts — track agent-referral traffic and AI-source revenue pre- and post-remediation over 60 days. Need a majority showing a directionally clear lift, and merchants need to be able to see it in their own analytics without a statistician.
3. Assumption: AI-drafted attributes are accurate enough that merchants approve in bulk rather than editing every row. How to test: run enrichment on 5 real catalogs across different categories, measure the accept-without-edit rate. Below ~70% and the review queue becomes worse than the original problem — merchants will abandon rather than approve 4,000 rows one at a time.
4. Assumption: The cold-audit email converts at 4–7% reply. How to test: send 300 pre-run-audit Looms before building the fix engine. This is cheap and fully falsifiable.
5. Assumption: Shopify doesn’t ship this natively within 12 months. How to test: unfalsifiable in advance — monitor Shopify Editions and dev changelogs monthly and treat it as a standing risk, not a validated assumption.

Risk flags

1. Platform dependency (severe). Shopify, OpenAI and Google each control both the spec and the surface. Shopify shipping native catalog-quality suggestions would take the Shopify segment to near-zero. Mitigation is real but partial: be multi-platform from early on and lean into WooCommerce, which has no platform patron and the worst data (24% schema adoption).
2. Commoditization (severe). Defensibility scores 2/10. Feed managers already own distribution to this exact buyer and can add enrichment as a feature. The counter is speed and focus — but this is an execution race, not a moat, and the business needs to be profitable at $1M–$2M ARR on the assumption that it never becomes defensible.
3. Accuracy and liability. A model that confidently invents “100% merino wool” for an acrylic blend creates a real consumer-protection problem for the merchant, not just a bad suggestion. This is why merchant-in-the-loop is non-negotiable, why enrichment must be grounded strictly in the merchant’s own images/copy/supplier docs, and why the system must mark low-confidence inferences rather than guessing. Get this wrong once publicly and the brand is finished.
4. Attribution ambiguity. AI-surface visibility is hard to measure cleanly. If merchants can’t see the lift in their own dashboard, they churn regardless of whether the product worked. Measurement is a product requirement, not a nice-to-have.
5. Market timing. Agent-driven transactions are still low single digits of total ecommerce. Bet is on the growth curve, not the current base — genuinely early, and if agentic commerce plateaus the urgency evaporates.

14. Structured verdict

Score:                  74/100
Verdict:                GO
Confidence:             Medium
Best-fit builder:       Technical solo founder or pair with ecommerce/feed-ops
                        background; must be comfortable publishing content weekly
Time to revenue:        6–8 weeks (remediation sprints pre-sold ahead of the
                        recurring product)
Capital to launch:      $8–12K (₹7–10L) — inference, hosting, prospect data
Top 3 assumptions to validate first:
  1. Catalog drift is continuous enough to justify recurring pricing — re-scan
     200 stores six weeks apart, count new defects per store
  2. Merchants accept AI-drafted attributes in bulk (>70% accept-without-edit)
     — run enrichment on 5 real catalogs across categories
  3. Pre-run-audit cold email replies at 4%+ — send 300 Looms before building
     the fix engine
Kill criteria:
  - Abandon if <70% of AI-drafted attributes are accepted without edit across
    5 test catalogs — the review burden then exceeds the value delivered
  - Abandon if median store adds <20 new catalog defects over six weeks — the
    recurring model is dead and this is a one-time services play
  - Abandon if Shopify ships native catalog-quality remediation, unless
    WooCommerce/BigCommerce alone can carry a $1M path
  - Abandon if <4% reply rate across 300 personalized audit Looms

15. Next step — 1-week validation sprint

• Day 1–2: Build the read-only scanner. Pull public product data and structured markup for 200 stores in the 500+ SKU band across Shopify and WooCommerce, score each against the published ACP and Google feed rules. Output: a defect distribution — which rules fail, how often, and whether the CommerceShop <1% figure reproduces on an independent sample.
• Day 3–4: Take the 20 worst-scoring stores. Run real enrichment on 100 SKUs each and hand-grade the output for accuracy against the merchant’s own images and copy. This produces the accept-rate number that assumption #3 lives or dies on — and it’s the number most likely to kill the idea, so it gets measured before anything is built.
• Day 5: Send 100 personalized Looms showing each merchant their three worst SKUs and what an agent sees instead. Offer a paid $1,500 remediation sprint, not a free trial.
• Decision rule (falsifiable): proceed only if (a) ≥85% of the 200 scanned stores fail at least one hard eligibility rule, (b) hand-graded enrichment accuracy is ≥70% on the 2,000 SKUs, and (c) ≥4 replies and ≥1 merchant agrees to pay for a sprint off 100 emails. Miss any one and the idea doesn’t get built — (b) failing means the product can’t work, (c) failing means nobody wants it bought.